Bulletin of the Seismological Society of America, Vol. 107, No. 2, pp. –, April 2017, doi: 10.

1785/0120160221

Ground-Motion Prediction Equation for the Chilean Subduction Zone

by Gonzalo A. Montalva, Nicolás Bastías,* and Adrian Rodriguez-Marek

Abstract A ground-motion prediction equation for the horizontal component of

the response spectral values from the Chilean subduction zone is developed. The data-
set contains 3774 recordings from 473 earthquakes, including the latest megathrust
events that occurred in the country (i.e., 2010 Mw 8.8 Maule, 2014 M w 8.1 Iquique,
and 2015 M w 8.3 Illapel). The functional form for the median model follows the pro-
posal by Abrahamson et al. (2016). Site effects are estimated based on V S30 . An addi-
tional model is built using only ground motions associated with measured values of
V S30 and reliable M w estimates (termed the high-quality [HQ] model). The standard
deviation of the HQ model is much lower than the main model, particularly the com-
ponent of the standard deviation that corresponds to site-to-site variability, indicating
that better site characterization can significantly reduce the overall uncertainty in
ground-motion estimation.

Electronic Supplement: Tables of regression coefficients for the high-quality

(HQ) model and 95% confidence intervals for model’s coefficients and MATLAB
implementation of the ground-motion prediction equation (GMPE).

Introduction
Chile is a country with high seismic activity, and most of 2006, for Japan; Lin and Lee, 2008, for Taiwan). A few
this activity is associated with the subduction of the Nazca GMPEs are available for the Chile subduction zone (Ruiz
plate beneath the South American plate. Subduction earth- and Saragoni, 2005; Boroscheck and Contreras, 2012), but
quakes have a strong impact on civil infrastructure and his- these models have theoretical shortcomings. For example,
torically caused economic and social losses. Two principal Boroschek and Contreras (2012) consider only 13 events in
types of seismicity can be associated with the subduction the dataset, and site information only differentiates rock from
mechanism: earthquakes occurring at the contact between soil, whereas Ruiz and Saragoni (2005) use epicentral dis-
the Nazca and South America plates (interface) and earth- tance as the distance measure and M s (surface-wave magni-
quakes occurring within the subducted oceanic plate and tude) as the magnitude measure. In fact, these two equations
related to tensional stresses within the plate (intraplate or would not be acceptable for modern PSHA studies, accord-
in-slab). Large-magnitude interface earthquakes that struck ing to the exclusion criteria proposed by Bommer et al.
Chile in recent years (e.g., 2010 M w 8.8 Maule, 2014 M w 8.1 (2010). The proposed GMPE for the Chilean subduction
Iquique, and 2015 Mw 8.3 Illapel) caused economic and life zone contributes to a better assessment of the seismic haz-
losses but have also allowed the scientific community to ard in the region and possibly to other subduction zones in
learn about these rarely recorded events. Among these lessons the world.
is the fact that the lack of a well-constrained model for these This article presents a GMPE for the Chilean subduction
types of events in this region hinders our ability to quantify zone based on data from events that occurred between 1985
and mitigate earthquake hazards. and 2015. Ground-motion records are selected and processed
Ground-motion prediction equations (GMPEs) are one with a uniform scheme and include interface and in-slab
of the essential components of probabilistic seismic-hazard events. The ground-motion parameters predicted by this
model are the peak ground acceleration (PGA, in g) and 5%
analysis (PSHA). A number of GMPEs have been proposed
damped pseudoacceleration response spectra (in g) up to a
for subduction environments based on worldwide data (e.g.,
period of 10 s.
Youngs et al., 1997; Atkinson and Boore, 2003, 2008;
Abrahamson et al., 2016) and regional data (e.g., Zhao et al.,
Strong Ground Motion Database
*Also at Camino a Chiguayante 5615, Department 404, Chiguayante The available strong-motion data in Chile has grown in
4100145, Chile; [email protected]. parallel with the growth of seismic networks, such as the

BSSA Early Edition / 1

2 G. A. Montalva, N. Bastías, and A. Rodriguez-Marek

multinational effort Integrated Plate boundary Observatory study (see Bastías and Montalva, 2016), whereas at other
Chile (hereafter, IPOC) and networks funded by the Chilean stations, published V S30 values are used. The number of
government only (e.g., Red Nacional de Acelerografos strong-motion stations with measured V S30 is limited (57
[RENADIC] and the Seismometer Network; both available stations, with 744 records); to obtain an inferred V S30 value,
by Centro Sismológico Nacional [CSN], see Data and Re- two methodologies are jointly applied (Bastías and Mon-
sources). The data used to fit the model presented herein are talva, 2016): the topographic slope of Wald and Allen
based on processed records from the flatfile published by (2007) and the use of the site’s predominant period as a
Bastías and Montalva (2015, 2016). The database includes proxy for V S30 (Zhao et al., 2006). A weighted average of
earthquake-related parameters (i.e., moment and local mag- these two methodologies is used. The numbers of stations
nitudes, epicentral coordinates, focal depth, and focal mecha- characterized following the approach stated above is 178.
nism), classification of earthquakes (i.e., interface, crustal, or The records were processed in a consistent manner,
in-slab), source-to-site measures (i.e., azimuth, epicentral, with each component individually filtered. The purpose of
hypocentral, and closest distances to rupture plane), and site the processing methodology is to identify the range of fre-
characterization at each recording station using average quencies for which the signal dominates over the noise or to
shear-wave velocity in the upper 30 m (V S30 ), topographic identify the time window of a signal in which noise is not
slope, and the site’s natural frequency (f0 ). The event meta- dominant. The procedure uses a band-pass filter in the fre-
data are gathered from public seismic catalogs (e.g., Global quency domain. To select the low-cut frequency, the S-wave
Centroid Moment Tensor [CMT]; International Seismologi- Fourier spectrum is divided by that of the noise window of
cal Centre [ISC]; CSN). Information about the instrument each record component. The ratio is then smoothed using a
type and stations coordinates was obtained from the net- Konno–Ohmachi filter to obtain the signal-to-noise ratio
work operators’ websites (IPOC, RENADIC, and CSN). (SNR). The low cutoff frequency for the filter is chosen at
The strong ground motion dataset uses first the hypocentral an SNR equal (or greater) to three, whereas the high cutoff
location (i.e., latitude, longitude, and depth) reported by the frequency is selected as the minimum value between the
CSN catalog and, if necessary, the location reported by the
Nyquist frequency and the frequency at which the spectrum
ISC catalog. Moment magnitudes and focal mechanisms
becomes flat (Akkar et al., 2011). The usable period range is
were obtained from the CMT catalog when available. How-
determined for each record as 1.25 times the low-cut corner
ever, when M w was unavailable, relationships between Mw
frequency used in the record processing, which is the same
and ML developed by Bastías and Montalva (2016) were
criterion used for processing the Next Generation Attenuation-
used.
West2 database (Ancheta et al., 2014). Within the processing,
Each earthquake was classified according to its location
we ensure that the level of noise present on each of the seismic
with respect to the trench axis (epicentral coordinates and
records is the same.
focal depth) and, when available, its focal mechanism. The
Part of the dataset, the high-quality (HQ) data (i.e., mea-
results of this classification are that, in general, interface
events are associated with reverse-faulting earthquakes and sured V S30 and moment magnitude reported by the Global
are located between the Perú–Chile trench and Chile’s coast CMT catalog) are used to create an alternative model that has
with focal depths up to 50 km, and in-slab events are asso- the potential for a better fit. Figure 1 shows both datasets:
ciated with normal-faulting earthquakes with focal depths entire data and the HQ subset.
greater than 50 km. Crustal events and subduction interface The initial data used to perform the regression include
were differentiated according to the event’s focal mecha- 3774 records from 473 earthquakes. This includes 2461
nism; a reverse faulting was associated with interface events. records from 281 interface events and 1313 records from
Earthquakes located between the Perú–Chile trench and 192 in-slab events. This large database allows an adequate
Chile’s coast, shallower than 50 km and associated with characterization of the attenuation parameters, magnitude
other-than-reverse mechanisms were cataloged as crustal scaling, and site response. The HQ subset includes 520 re-
events. For events without focal mechanism information, we cords from 208 earthquakes, including 411 records from
used a slab subduction model (Hayes et al., 2012) for their 151 interface events and 109 records from 57 in-slab events.
classification. In this case, events for which the hypocenter Figure 1 shows the distributions of moment magnitude
location matched the location of the interface were classified against source-to-site distance and against focal depth, seg-
as interface events. The rupture plane for the most significant regated by type of seismogenesis (i.e., interplate or in-slab).
events is obtained from the scientific literature. For earth- The strong ground motion data were obtained from 235 sta-
quakes without a published rupture plane, the dimensions tions with a V S30 range between 108 and 1951 m=s. The
of the rupture are estimated using the empirical relationships geometric mean of the horizontal components, for the 5%
proposed by Strasser et al. (2010), and finite-fault distances damped spectral acceleration, is computed for PGA, and 22
are computed by assuming that the centroid reported by spectral periods between 0.02 and 10 s, using time-domain
Global CMT is at the center of the rupture plane. The site integration. The spectral accelerations and metadata used
conditions at ground-motion stations were parameterized are freely available from the NEEShub platform (Bastías
using V S30 . For some stations, V S30 is measured within this and Montalva, 2015).
BSSA Early Edition
Ground-Motion Prediction Equation for the Chilean Subduction Zone 3

EQ-TARGET;temp:intralink-;df5;313;721 fevent=depth  θ10  θ11
min
Zh ; 120 − 60Fevent
5

fsite
PGA1000 ; V S30 

EQ-TARGET;temp:intralink-;df6;313;697

8  
V
> θ12 ln V linS − b ln
PGA1000  c
>
>
>
<    n 
V
  b ln PGA1000  c V linS ; if V S30 < V lin
>
>
>
>    
: V V
θ12 ln V linS  b ln V linS ; if V S30 < V lin

6

1000; if V S30 > 1000
V S 
7
V S30 ; if V S30 ≤ 1000
EQ-TARGET;temp:intralink-;df7;313;567

8  
>
>
< θ7 θ8 ln ×FFABA ; ifFevent 1
max
R;85
40
fFABA
R    ;
>
>
EQ-TARGET;temp:intralink-;;313;526

: θ15 θ16 ln max
R;100

40 ×F FABA ; ifF event 0
Figure 1. Distribution of records and earthquakes from the
strong ground motion database used for the regression. (a,b) Full
dataset; (c,d) high-quality (HQ) dataset. The color version of this in which μ
Zjθ is the median model, with Z being the matrix
figure is available only in the electronic edition. that contains all explicative variables and θ being the coef-
ficients of the regression; SA is the 5% damped spectral ac-
Functional Form and Regression Methodology celeration or PGA in units of g; M w is the moment magnitude
of the earthquake; Zh is the hypocentral depth in kilometers;
We selected the functional form by Abrahamson et al. R is the source-to-site distance, which is defined as the
(2016) developed for the BC Hydro project. The selection of closest distance to the rupture plane (Rrup ) for interface earth-
the BC Hydro functional form is due to the good fit to a sub- quakes and the hypocentral distance for in-slab earthquakes
set of the data from the Chilean subduction zone used in this (Rhyp ); PGA1000 is the median PGA value for a site with V S30
study (Bastías et al., 2015). This functional form has theo- equal to 1000 m=s; Fevent is a dummy variable that takes the
retical advantages over others because it uses V S30 as site value 1 when the record is from an in-slab earthquake and
explanatory variable and includes nonlinear site response. 0 for interface earthquakes. The fFABA term represents the
The median model is described in the following equations: fore-arc-to-back-arc model scaling, in which fFABA is a
binary variable equal to 1 for back-arc sites and 0 for fore-
μ
Zjθ  ln SA
T
EQ-TARGET;temp:intralink-;df1;55;291

arc or unknown sites. The values of coefficients ΔC1 , θ9 , C4 ,

 θ1  fsource  fpath  f event=depth  fsite  fFABA V lin , b, c, and n are adopted directly from the BC Hydro
model, as is the back-arc effect. These were kept because

1 there are no data from back-arc sites and no sufficient data
for the nonlinear site effects.
Another modification of the BC Hydro equation is that in
EQ-TARGET;temp:intralink-;df2;55;211 fsource  θ4 ΔC1  fmag
Mw 
2 its original form θ3 –θ5 are fitted for PGA and kept constant for
all other spectral periods; in this work, these are fitted for each

period. Also, the submodel associated with magnitude scaling
θ4
M w −
C1  ΔC1 ; if M w ≤ C1  ΔC1 is modified to fit a bilinear behavior (excluding coefficient
fmag
Mw  
θ5
M w −
C1  ΔC1 ; if M w > C1  ΔC1 θ13 from fmag ). An important assumption of the model is that
EQ-TARGET;temp:intralink-;df3;55;189

3 the nonlinear site behavior should apply to subduction earth-

quakes in the same way that they apply to crustal events
(Walling et al., 2008; Abrahamson et al., 2016).
To fit the empirical data, we use a nonlinear mixed ef-
EQ-TARGET;temp:intralink-;df4;55;115

fpath  θ2  θ14 Fevent  θ3
Mw − 7:8 fects regression. Stafford (2014) demonstrated that this meth-
odology has advantages over methods that split the residuals
× ln
R  C4 exp
θ9
M w − 6  θ6 R
4 in a multistage scheme, because failure to include all random
BSSA Early Edition
4 G. A. Montalva, N. Bastías, and A. Rodriguez-Marek

three or more records, and nearly 60% of stations have three

or more records. The full dataset has a maximum number of
records used in this regression of 3717, and the HQ dataset
has 515 usable records. The final regression forced the co-
efficient θ6 to negative values to avoid unrealistic behavior of
distance attenuation. The number of records used for each
period is presented in Table 1 and Ⓔ Table S1 (available in
the electronic supplement to this article), for the full dataset
and the HQ dataset, respectively.

Ground-Motion Prediction Model

The values of the coefficients for the median spectral
acceleration and standard deviations are presented in Table 1.
The central value of ΔC1 recommended by Abrahamson
et al. (2016) is kept in the proposed model. Table 1 includes
Figure 2. Number of data used in the regression segregated by the terms that are kept from the BC Hydro model as well.
quality of data. The color version of this figure is available only in
Figures 3 and 4 show the mean response spectra predicted
the electronic edition.
by the proposed model for a site on the fore-arc, with V S30 
300 m=s (National Earthquake Hazards Reduction Program
effects directly in the regression may produce a bias in the
[NEHRP] class D). A MATLAB script for the Chile GMPE
median of the model. The lme4 package of the statistical
is available in the Ⓔ electronic supplement. The regression
software R developed by Bates et al. (2015) provides an ef-
coefficients for the HQ model are provided in Ⓔ Table S1.
ficient computational method to manage models fitted using
a nonlinear mixed-effects regression. The total difference
(i.e., residual) between the natural log of the observed spec- Sensitivity Analysis of Regression Coefficients
tral acceleration, ln
SAobs , and the median prediction of the
To define the 95% confidence intervals for all the model
model (i.e., μ
Zjθ is divided into three components. The
coefficients, a bootstrap analysis is performed. 1000 boot-
average residual for each earthquake e
δBe , the average
strap replications, using datasets with the same number of
event-corrected residual at each station (δS2SS ), and the re-
records used in the original database, but accepting duplicate
maining residual, or single-station residual (δW 0 ). This is
data, allow the same number of regressions. The confidence
expressed as
intervals are obtained through the bias-corrected percentile
method (Efron and Tibshirani, 1994); these intervals are sum-
EQ-TARGET;temp:intralink-;df8;55;355 ln
SAobs   μ
Zjθ  δBe  δS2SS  δW 0 ;
8 marized in Ⓔ Table S2. The same intervals for the HQ model
are listed in Ⓔ Table S3, which reflect the higher uncertainty
in which the total residual is decomposed into a site-to-site on the coefficients for this model. This higher uncertainty may
normally distributed residual term with standard deviation be attributed to the lower number of records.
ϕS2S (δS2Ss ∼ N
0; ϕ2S2S ), a between-earthquake residual In addition, a resampling analysis is performed to evalu-
term (δBe ∼ N
0; τ2 ), and single-station (δW 0 ∼ N
0; ϕ2SS ) ate the sensitivity of the model’s coefficients to the number of
residual component. The total standard deviation of the records in the dataset. Several subsets were created with dif-
model is given by the square root of the sum of their squares ferent sample sizes from 500 to 3500 records with 500 re-
p
(i.e., σ T  τ2  ϕ2S2S  ϕ2SS ). cords intervals; these records were randomly selected. For
A first run of the regression is performed with all data each sample size, 100 regressions were performed to obtain
from the Chile strong ground motion database. The Rosner a distribution of each coefficient for each sample size. The
(1975, 1983) algorithm is then used to identify outliers in the resampling analysis allows the study of the convergence of
intensities for each period. After the removal of the outliers, a statistical tests (e.g., log-likelihood [LLH], p-value) to evalu-
new regression is performed to define the coefficients of the ate ground-motion prediction models, similar to the analysis
final model. The model considers no smoothing of coeffi- by Azarbakht et al. (2014). The random selection of records,
cients and deviations, and this task can be done by the hazard within the resampling analysis, implies that a small sample
analyst if deemed necessary (Akkar and Bommer, 2010). The size (e.g., 500 records) can lack records of an entire earth-
filtering of the data and the removal of outliers imply that the quake. A large sample size (e.g., 3500) has almost the entire
number of data used to fit the model changes with the period database included, and this is because repetition is not al-
and with the catalog (i.e., full dataset and HQ). The number lowed (i.e., no record can be selected twice). If the values of
of usable records is shown in Figure 2. No exclusion based the coefficients are strongly affected by a small change in the
on the number of records per station or per event was used to size of the ground-motion dataset, then the predicted values
select the final dataset. However, about 70% of events have may not be reliable, and the uncertainty would be high. The
BSSA Early Edition
 Table 1
Model Coefficients
Number of
Period (s) Data Points θ1 θ4 θ5 θ2 θ3 θ6 θ14 θ10 θ11 θ12 τ ϕS2S ϕSS

0.00 3657 5.87504 0.80277 −0.33487 −1.75360 0.13125 −0.00039 −0.73080 4.53143 0.00567 1.01495 0.47462 0.56436 0.39903
0.02 3657 5.97631 0.84132 −0.28055 −1.77011 0.12246 −0.00039 −0.73869 4.57416 0.00565 1.03738 0.47632 0.57188 0.40261
0.05 3658 7.45297 1.03131 −0.03954 −2.03336 0.08332 0.00000 −0.69849 4.56071 0.00848 1.31034 0.53776 0.57850 0.39720
0.075 3656 8.04760 1.03437 −0.01295 −2.10610 0.08013 −0.00010 −0.65336 4.36639 0.00922 1.48158 0.56188 0.59937 0.38824
0.1 3652 7.76085 1.07565 0.00758 −1.99371 0.07303 −0.00079 −0.55051 3.90923 0.00630 1.65619 0.52707 0.63410 0.38365
0.15 3654 6.17192 1.17061 0.10491 −1.58654 0.05482 −0.00268 −0.42997 3.06236 0.00559 1.93944 0.50642 0.63022 0.39930
0.2 3657 4.83403 1.20531 0.17968 −1.29711 0.05250 −0.00338 −0.53088 3.50113 0.00320 2.08901 0.44619 0.61699 0.41782
0.25 3682 4.42688 1.37607 0.22912 −1.18774 0.02995 −0.00355 −0.58086 3.62816 0.00182 2.25003 0.45040 0.58609 0.43277
0.3 3673 4.57009 1.34991 0.15593 −1.24896 0.03866 −0.00245 −0.66281 3.87634 0.00213 2.28339 0.42549 0.57014 0.44123
0.4 3643 3.98311 1.37954 0.11671 −1.13377 0.04683 −0.00208 −0.72244 4.03388 0.00069 2.31409 0.42945 0.54796 0.45157
0.5 3591 4.86034 1.51950 0.18348 −1.38020 0.03822 −0.00002 −0.79644 4.31418 0.00065 2.33333 0.43334 0.49113 0.45476
0.6 3634 4.67510 1.66663 0.21968 −1.35362 0.02524 0.00000 −0.90120 4.75197 0.00087 2.23422 0.44599 0.49078 0.45219
0.75 3614 4.30862 1.85625 0.29783 −1.30800 0.00995 0.00000 −0.89829 4.70452 −0.00031 2.05217 0.46723 0.48213 0.45553
1 3685 3.57339 1.81217 0.24372 −1.23082 0.03605 0.00000 −0.87331 4.56020 −0.00101 1.63506 0.50143 0.45955 0.43828
1.5 3717 2.92216 2.03469 0.22521 −1.18750 0.02769 −0.00010 −0.94686 4.83343 0.00010 0.69338 0.51633 0.42573 0.42297
2 3648 2.39780 2.04340 0.27383 −1.16319 0.04011 −0.00033 −0.90845 4.59029 0.00109 −0.09762 0.50688 0.40179 0.40377
2.5 3583 1.64148 1.88987 0.18740 −1.06544 0.08310 −0.00121 −0.80518 4.13415 0.00035 −0.34932 0.51465 0.39825 0.38489
3 3525 1.66483 1.90504 0.13268 −1.12678 0.09404 −0.00088 −0.81689 4.18978 0.00073 −0.33270 0.50365 0.38493 0.37384
4 3283 0.90565 1.71178 0.01380 −1.07620 0.13838 −0.00062 −0.87331 4.50907 0.00084 −0.41321 0.45311 0.35579 0.37020
5 3102 0.61234 1.59359 0.06465 −1.13080 0.15259 0.00000 −0.87800 4.56386 0.00068 −0.42395 0.43900 0.34991 0.37674
6 2921 0.32672 1.69184 0.32368 −1.15734 0.12421 0.00000 −0.88436 4.55837 0.00137 −0.38760 0.42084 0.32048 0.37173
7.5 2780 −0.24140 1.71126 0.60252 −1.14070 0.10951 0.00000 −0.98803 5.08282 0.00167 −0.32638 0.41701 0.29895 0.36349
10 2473 −0.96314 1.67160 0.77621 −1.09295 0.11344 0.00000 −1.05008 5.49692 −0.00070 −0.25811 0.38872 0.28454 0.36173
Ground-Motion Prediction Equation for the Chilean Subduction Zone

BSSA Early Edition

(continued)
5
 6

Table 1 (Continued)
Period (s) σ θ7 θ8 θ15 θ16 θ9 ΔC1;interface ΔC1;in-slab V lin b n c C4 C1

0.00 0.83845 1.0988 −1.420 0.9969 −1.000 0.4 0.200 −0.300 865.1 −1.186 1.18 1.88 10 7.2
0.02 0.84618 1.0988 −1.420 0.9969 −1.000 0.4 0.200 −0.300 865.1 −1.186 1.18 1.88 10 7.2
0.05 0.88409 1.2536 −1.650 1.1030 −1.180 0.4 0.200 −0.300 1053.5 −1.346 1.18 1.88 10 7.2
0.075 0.90867 1.4175 −1.800 1.2732 −1.360 0.4 0.200 −0.300 1085.7 −1.471 1.18 1.88 10 7.2
0.1 0.90944 1.3997 −1.800 1.3042 −1.360 0.4 0.200 −0.300 1032.5 −1.624 1.18 1.88 10 7.2
0.15 0.90171 1.3582 −1.690 1.2600 −1.300 0.4 0.200 −0.300 877.6 −1.931 1.18 1.88 10 7.2
0.2 0.86853 1.1648 −1.490 1.2230 −1.250 0.4 0.200 −0.300 748.2 −2.188 1.18 1.88 10 7.2
0.25 0.85654 0.9940 −1.300 1.1600 −1.170 0.4 0.200 −0.300 654.3 −2.381 1.18 1.88 10 7.2
0.3 0.83713 0.8821 −1.180 1.0500 −1.060 0.4 0.200 −0.300 587.1 −2.518 1.18 1.88 10 7.2
0.4 0.82982 0.7046 −0.980 0.8000 −0.780 0.4 0.144 −0.300 503 −2.657 1.18 1.88 10 7.2
0.5 0.79737 0.5799 −0.820 0.6620 −0.620 0.4 0.100 −0.300 456.6 −2.669 1.18 1.88 10 7.2
0.6 0.80265 0.5021 −0.700 0.5800 −0.500 0.4 0.074 −0.300 430.3 −2.599 1.18 1.88 10 7.2
0.75 0.81134 0.3687 −0.540 0.4800 −0.340 0.4 0.042 −0.300 410.5 −2.401 1.18 1.88 10 7.2
1 0.80914 0.1746 −0.340 0.3300 −0.140 0.4 0.000 −0.300 400 −1.955 1.18 1.88 10 7.2
1.5 0.79168 −0.0820 −0.050 0.3100 0.000 0.4 −0.058 −0.300 400 −1.025 1.18 1.88 10 7.2
2 0.76249 −0.2821 0.120 0.3000 0.000 0.4 −0.100 −0.300 400 −0.299 1.18 1.88 10 7.2
2.5 0.75605 −0.4108 0.250 0.3000 0.000 0.4 −0.155 −0.300 400 0 1.18 1.88 10 7.2
3 0.73593 −0.4466 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2
4 0.68480 −0.4344 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2
5 0.67609 −0.4368 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2
6 0.64653 −0.4586 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2
7.5 0.62881 −0.4433 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2
10 0.60243 −0.4828 0.300 0.3000 0.000 0.4 −0.200 −0.300 400 0 1.18 1.88 10 7.2

BSSA Early Edition

G. A. Montalva, N. Bastías, and A. Rodriguez-Marek
Ground-Motion Prediction Equation for the Chilean Subduction Zone 7

Figure 3. Response spectra obtained for a fore-arc site with V S30  300 m=s for an interplate earthquake. PGA, peak ground accel-
eration. The color version of this figure is available only in the electronic edition.

Figure 4. Response spectrum obtained for a fore-arc site with V S30  300 m=s for an in-slab earthquake. The color version of this figure
is available only in the electronic edition.

BSSA Early Edition

8 G. A. Montalva, N. Bastías, and A. Rodriguez-Marek

Figure 5. Resampling analysis to evaluate the sensitivity of coefficients θ1 and θ2 for PGA and 0.1 s. Dashed horizontal lines are the
confidence intervals for each coefficient from the bootstrap analysis. (a) and (c) correspond to PGA, and (b) and (d) correspond to spectral
acceleration (SA) (T  0:1 s). The color version of this figure is available only in the electronic edition.

convergence of each coefficient occurs at subsets with at V S30 values. The probability of exceeding any specific inten-
least 1000 records, but the mean for even the smallest sample sity value decreases with a decrease in sigma, so another ad-
(500 records) is within the 95% confidence interval from the vantage of this alternative model is its lower variability; this
full dataset (i.e., from the bootstrap analysis). Figure 5 shows comes at the cost of better input data (i.e., measured V S30 ), and
this exercise for coefficients θ1 and θ2 , for which the box- a restricted usable distance range (Rrup < 300 km).
plots show the mean ± one standard deviation for each sam- The best-characterized event in the dataset is the 2010
ple size. With larger sample sizes, the standard deviation for Mw 8.8 Maule earthquake, because many of the recording
each coefficient decreases, and in fact it would eventually sites have a measured V S30 value (e.g., Kayen et al., 2014;
reach zero with the coefficient taking the model’s value. The Bastías and Montalva, 2016). Figure 7 overlays the records
same behavior occurs for the remaining coefficients. from this event with the distance attenuation for both models
for an M w 8.8 earthquake and a V S30 value close to the
Residual and Uncertainty Analysis median of those stations that recorded the Maule event (the
median is 515 m=s). The standard deviations for the HQ
The accuracy of the model with respect to the empirical model are listed in Table 2.
data is examined by plotting the independent variables against To assess the difference in goodness of fit between the
the residuals. Figure 6 shows the between-event residual Abrahamson et al. (2016) and the models developed in this
against moment magnitude, the single-station residual against study, we compare the performance of each model using the
source-to-site distance, and the site-to-site residual versus LLH methodology proposed by Scherbaum et al. (2009).
V S30 . Mean and standard deviations of binned data are com- This allows a comparison, under an objective scheme, of the
puted and shown in the figure. The tick marks on the x axis predictive capability of each model for this database. Figure 8
define the bin range values used for analysis. After testing all shows the LLH values for the HQ dataset, the Abrahamson
independent variables, no trends are found. Consequently, the et al. (2016) model, and the ones proposed in this study.
residual plots lead us to conclude that the regression procedure Models developed in this article show better performance
is robust and reliable. (i.e., lower LLH values). Ⓔ Figure S1 shows a comparison
For engineering applications in seismic-hazard analyses, in terms of distance attenuation.
the standard deviations are as important as the median pre-
dictions. The HQ model, fitted only with HQ data, allows the Comparison of Standard Deviations with Other
assessment of the uncertainty in the input variables. The goal Tectonic Regions
of this model is to remove some of the scatter introduced
by the uncertainty associated with the conversion of local Rodriguez-Marek et al. (2013) compared ergodic
p
magnitudes to moment magnitudes and from the inferred within-event standard deviations (ϕ with ϕ  ϕ2SS  ϕ2S2S )
BSSA Early Edition
Ground-Motion Prediction Equation for the Chilean Subduction Zone 9

Figure 6. Plots of residual components versus predictive variables. Rows one through four correspond to PGA, SA (T  0:1 s),
SA (T  1 s), and SA (T  10 s). The color version of this figure is available only in the electronic edition.

and event-corrected single-station standard deviations (ϕSS ) earthquake); consequently, it is feasible to extend the appli-
for multiple regions, concluding that in particular ϕSS is a cability of the model up to Mw 9. For V S30, the range of
very stable parameter across different tectonic environments. validity goes from 100 to 1000 m=s. In this article, we also
The same data are used to compare the standard deviations of regress for a GMPE using only a HQ subset of the database
the proposed model. Figure 9a shows remarkable agreement (i.e., the HQ model). The validity of the HQ model is re-
of the proposed model with the parameters obtained from stricted to interface events and distances up to 300 km. Both
other regions and tectonic environments. The HQ model has models predict similar median response over the range of
significantly lower ϕS2S values due to the better-characterized applicability (see Fig. 7), but the HQ model renders signifi-
sites, almost identical τ, and slightly lower ϕSS values. The val- cantly lower values of ϕS2S . This reduction is likely due to
ues for ϕ are also within the parameters obtained worldwide. the better metadata (V S30 ) for the HQ dataset. Other studies
(Rodriguez-Marek et al., 2013) indicated decreasing ϕSS val-
Summary and Conclusions ues for large magnitudes, which may indicate why the HQ
This article presented a GMPE for Chilean earthquakes model, which has a much larger proportion of large M w
based on a comprehensive database compiled by Bastías and events on its dataset, has a slightly lower ϕSS.
Montalva (2016). Using the criteria proposed by Bommer The difference between the total intra-event variability
et al. (2010), the range of validity of the proposed GMPE (ϕ) and the single-station event-corrected variability (ϕSS )
is for distances up to 300 km; however, distances up to is a clear indication that a better constraint on site parameters
1000 km show reasonable behavior. For magnitudes, the reduces the value of ϕS2S , which in turn implies a reduction
range goes from 5.0 to 8.0, in which the bulk of the data lies. in ϕ. These findings are also true for other regions (e.g.,
However, events in the Mw 8.0–9.0 bin are some of the Thompson and Wald, 2016) in which improved site terms
best-characterized events (including the M w 8.8 Maule analysis within the GMPE is needed.

BSSA Early Edition

10 G. A. Montalva, N. Bastías, and A. Rodriguez-Marek

Table 2 Data and Resources

High-Quality Model Standard Deviations
All the data used in this work are freely available at
Period (s) ϕSS ϕS2S τ
https://nees.org/resources/13694 (last accessed January
0.00 0.29315 0.35438 0.48274 2017), which has been processed according to Bastías and
0.02 0.29672 0.35919 0.49080
0.05 0.29003 0.38169 0.53613 Montalva (2016). Data sources include Centro Sismologico
0.075 0.27892 0.41023 0.58232 Nacional (CSN, National Seismological Center) available at
0.1 0.28010 0.45716 0.56901 http://evtdb.csn.uchile.cl/ (last accessed January 2017), Inte-
0.15 0.29938 0.42642 0.53998 grated Plate boundary Observatory Chile (IPOC), and from
0.2 0.33550 0.39664 0.46322
0.25 0.32956 0.36564 0.46462 RENADIC from the University of Chile available at http://
0.3 0.33449 0.35375 0.41917 www.renadic.cl/ (last accessed January 2017). For the regres-
0.4 0.36362 0.37225 0.41862 sions, the software R (http://www.R-project.org/, last ac-
0.5 0.37367 0.31389 0.43451 cessed January 2017) was used.
0.6 0.36857 0.29805 0.43404
0.75 0.36698 0.33310 0.45313
1 0.38278 0.36033 0.46525
1.5 0.36307 0.40814 0.42219
2 0.37107 0.37199 0.38592
2.5 0.36636 0.36573 0.38407
3 0.33519 0.39403 0.38146
4 0.33086 0.39148 0.38539
5 0.31169 0.40978 0.40439
6 0.30916 0.38271 0.39921
7.5 0.29010 0.21232 0.41293
10 0.29420 0.28240 0.40111

Figure 8. Log-likelihood (LLH) values for Chilean models and

Abrahamson et al. (2016) ground-motion prediction equation
model for HQ dataset. The color version of this figure is available
only in the electronic edition.

Figure 7. Comparison between models fitted with all data and

with HQ data. Main model mean and standard deviation are in
black, and HQ model is in gray. The Maule earthquake data are
plotted on top of the curves, where solid points are well-character-
ized stations (HQ data). The curves show attenuation with distance Figure 9. Comparison of standard deviation models from other
for M w 8.8 and V S30 value of 500 (m=s). The color version of this tectonic regions and the results of this study. The color version of
figure is available only in the electronic edition. this figure is available only in the electronic edition.

BSSA Early Edition

Ground-Motion Prediction Equation for the Chilean Subduction Zone 11

Acknowledgments Lin, P.-S., and C.-T. Lee (2008). Ground-motion attenuation relationships
for subduction-zone earthquakes in northeastern Taiwan, Bull. Seis-
This work was partially funded by Fondo de Fomento al Desarrollo mol. Soc. Am. 98, no. 1, 220–240.
Científico y Tecnológico (FONDEF) D10E1027 and Comisión Nacional de Rodriguez-Marek, A., F. Cotton, N. A. Abrahamson, S. Akkar, L. Al Atik,
Investigación Científica y Tecnológica/Fondo de Financiamiento de Centros B. Edwards, G. A. Montalva, and H. M. Dawood (2013). A model for
de Investigación en Áreas Prioritarias (CONICYT/FONDAP/15130015), and single-station standard deviation using data from various tectonic
this support is greatly acknowledged. We thank the three anonymous reviewers regions, Bull. Seismol. Soc. Am. 103, 3149–3163.
and Associate Editor John Douglas for their careful and detailed reviews. Rosner, B. (1975). On the detection of many outliers, Technometrics 17,
221–227.
Rosner, B. (1983). Percentage points for a generalized ESD many-outlier
References
procedure, Technometrics 25, 165–172.
Abrahamson, N., N. Gregor, and K. Addo (2016). BC Hydro ground motion Ruiz, S., and G. Saragoni (2005). Attenuation equations for subduction-zone
prediction equations for subduction earthquakes, Earthq. Spectra 32, earthquakes in Chile considering two seismogenic mechanisms and
no. 1, 23–44. site effects, IX Jornadas Chilenas de Sismología e Ingeniería Antisís-
Akkar, S., and J. J. Bommer (2010). Empirical equations for the prediction mica, Concepción, Chile, 16–19 November 2005.
of PGA, PGV, and spectral acceleration in Europe, the Mediterranean Scherbaum, F., E. Delavaud, and C. Riggelsen (2009). Model selection in
region, and the Middle East. Seismol. Res. Lett. 81, no. 2, 195–206. seismic hazard analysis: An information-theoretic perspective, Bull.
Akkar, S., Ö. Kale, E. Yenier, and J. Bommer (2011). The high-frequency Seismol. Soc. Am. 99, 3234–3247.
limit of usable response spectral ordinates from filtered analogue and Stafford, P. J. (2014). Crossed and nested mixed-effects approaches for
digital strong-motion accelerograms, Earthq. Eng. Struct. Dynam. 40, enhanced model development and removal of the ergodic assumption
1387–1401. in empirical ground-motion models, Bull. Seismol. Soc. Am. 104,
Ancheta, T. D., R. B. Darragh, J. P. Stewart, E. Seyhan, W. J. Silva, B. S.-J. 702–719.
Chiou, K. E. Wooddell, R. W. Graves, A. R. Kottke, D. M. Boore, Strasser, F. O., M. C. Arango, and J. J. Bommer (2010). Scaling of the source
et al. (2014). NGA-West2 database, Earthq. Spectra 30, no. 3, dimensions of interface and intraslab subduction-zone earthquakes
989–1005. with moment magnitude, Seismol. Res. Lett. 81, 941–950.
Atkinson, G. M., and D. M. Boore (2003). Empirical ground-motion rela- Thompson, E. M., and D. J. Wald (2016). Uncertainty in V S30 -based site
tions for subduction-zone earthquakes and their application to Cascadia response, Bull. Seismol. Soc. Am. 106, no. 2, 453–463.
and other regions, Bull. Seismol. Soc. Am. 93, 1703–1729. Wald, D. J., and T. I. Allen (2007). Topographic slope as a proxy for seismic
Atkinson, G. M., and D. M. Boore (2008). Erratum to empirical ground- site conditions and amplification, Bull. Seismol. Soc. Am. 97, 1379–1395.
motion relations for subduction-zone earthquakes and their application Walling, M., W. Silva, and N. A. Abrahamson (2008). Nonlinear site am-
to Cascadia and other regions, Bull. Seismol. Soc. Am. 98, 2567–2569. plification factors for constraining the NGA models, Earthq. Spectra
Azarbakht, A., S. Rahpeyma, and M. Mousavi (2014). A new methodology 24, 243–255.
for assessment of the stability of ground-motion prediction equations, Youngs, R. R., S. J. Chiou, W. J. Silva, and J. R. Humphrey (1997). Strong
Bull. Seismol. Soc. Am. 104, 1447–1457. ground motion attenuation relationships for subduction zone earth-
Bastías, N., and G. A. Montalva (2015). Chile Strong Ground Motion Flat- quakes, Seismol. Res. Lett. 68, 58–73.
file, NEEShub website, available at https://nees.org/resources/13694 Zhao, J. X., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi, H. Ogawa,
(last accessed January 2017). K. Irikura, H. K. Thio, P. G. Somerville, et al. (2006). Attenuation rela-
Bastías, N., and G. A. Montalva (2016). Chile strong ground motion flatfile, tions of strong ground motion in Japan using site classification based on
Earthq. Spectra 32, no. 4, 2549–2566, doi: 10.1193/102715EQS158DP. predominant period, Bull. Seismol. Soc. Am. 96, 898–913.
Bastías, N., G. A. Montalva, F. Leyton, E. Saez, F. Ruz, and P. Troncoso
(2015). Evaluation of ground motion prediction equations (GMPEs)
for Chile subduction zone, 15th Pan-American Conf. on Soil Mechanics
and Geotechnical Engineering, Buenos Aires, Argentina, November. Departamento de Ingeniería Civil
Bates, D., M. Maechler, B. Bolker, and S. Walker (2015). Fitting linear Universidad de Concepción
mixed-effects models using lme4, J. Stat. Software 67, no. 1, 1–48. Edmundo Larenas 219, Casilla 160-C, Correo 3
Bommer, J. J., J. Douglas, F. Scherbaum, F. Cotton, H. Bungum, and D. Fäh Concepción 4070409
Chile
(2010). On the selection of ground-motion prediction equations for
[email protected]
seismic hazard analysis, Seismol. Res. Lett. 81, no. 5, 783–793. [email protected]
Boroschek, R., and V. Contreras (2012). Strong ground motion from the (G.A.M., N.B.)
2010 Mw 8.8 Maule Chile earthquake and attenuation relations for
Chilean subduction zone interface earthquakes, Proc. International
Symposium on Engineering Lessons Learned from the 2011 Great East
Japan Earthquake, Tokyo, Japan, 1–4 March. The Charles Edward Via
Efron, B., and R. J. Tibshirani (1994). An Introduction to the Bootstrap, Chap- Jr. Department of Civil and Environmental Engineering
man and Hall/CRC, New York, New York, ISBN: 9780412042317. Virginia Tech, 200 Patton Hall
Hayes, G. P., D. J. Wald, and R. L. Johnson (2012). Slab1.0: A three-dimensional Blacksburg, Virginia 24061
model of global subduction zone geometries, J. Geophys. Res. 117, [email protected]
no. B01302, doi: 10.1029/2011JB008524. (A.R.-M.)
Kayen, R., B. D. Carkin, S. Corbet, C. Pinilla, A. Ng, E. Gorbis, and C. Truong
(2014). Seismic velocity site characterization of thirty-one Chilean seis-
mometer stations by spectral analysis of surface wave dispersion, Pacific Manuscript received 8 July 2016;
Earthquake Engineering Research Reports, No. 2014/05. Published Online 7 February 2017

BSSA Early Edition